AWS publishes reference build for voice airline concierge using Nova 2.5 Sonic and Bedrock AgentCore

AWS has published a reference implementation for a voice travel concierge that runs Amazon Nova 2.5 Sonic, a speech-to-speech model, on Amazon Bedrock AgentCore runtime. The agent is built with the Strands Agents framework and reaches backend systems through AgentCore Gateway using the Model Context Protocol (MCP). The code is in the aws-samples GitHub repository.

This is a tutorial and architecture guide, not a model launch. The post does not disclose pricing, context window, parameter count, or benchmark scores for Nova 2.5 Sonic. Pricing is not yet disclosed in this source.

What the concierge does

According to AWS, a traveler can speak to retrieve an itinerary, change a seat, update a meal preference, ask a policy question, or request a live agent. The voice layer runs alongside existing app screens, so users can switch between tapping and talking in one session.

Architecture

AWS splits the solution into four sections, deployed through AWS CDK:

• Backend (5 CDK stacks): DynamoDB tables, Lambda functions, API Gateway REST endpoints with IAM authorization, Bedrock Knowledge Bases, and Amazon Cognito.
• AgentCore Gateway (1 stack): Exposes each backend endpoint as an MCP tool the agent can call by name.
• AgentCore runtime (2 stacks): ECR stores the container image, S3 holds source uploads, and CodeBuild produces an ARM64 Docker image. The runtime supports WebSocket connections and provides microVM isolation per session.
• Front end (1 stack): A React app hosted on AWS Amplify.

Request flow

1. The user logs in; Cognito returns JWTs and temporary AWS credentials.
2. The front end opens a SigV4-signed WebSocket to AgentCore runtime.
3. The runtime validates the token and initializes Nova 2.5 Sonic through Bedrock.
4. Nova 2.5 Sonic processes speech and triggers tool calls. The agent calls AgentCore Gateway via MCP, which forwards to API Gateway and then Lambda, which queries DynamoDB.
5. The synthesized voice response streams back over the same WebSocket.

For policy questions, the Gateway queries Bedrock Knowledge Bases directly through an MCP Knowledge Base connector. The knowledge base returns policy passages as citations. Per AWS, it uses a service-managed embedding model, so no separate embedding model access is required.

For live-agent handoff, a Lambda function logs the escalation in DynamoDB and returns a reference number. Amplify then triggers a call from the user's device. The post's summary also mentions an estimated wait time. Amazon SES sends email notifications, CloudWatch collects logs and metrics, and AWS KMS encrypts data at rest.

Prerequisites

• An AWS account with Bedrock model access for Nova 2.5 Sonic in the deployment Region (availability varies by Region)
• Node.js 20.x or later
• Python 3.12 or later for data seeding and the test client
• AWS CLI 2.x and AWS CDK CLI 2.x, with the account bootstrapped via npx cdk bootstrap

Deployment runs from a single script. The sample backend uses synthetic data covering itineraries, seat maps, passenger updates, flight status, loyalty, policy lookups, and escalation.

What this means

The most useful detail is the loose coupling. Because the agent reaches backends only through MCP tools on AgentCore Gateway, an airline can wrap existing REST APIs without rewriting them for the voice agent. The same pattern applies outside travel.

The post also shows AWS positioning Nova 2.5 Sonic inside its agent stack, not as a standalone model. Teams evaluating speech-to-speech options will find no latency figures, pricing, or quality benchmarks here, so cost and responsiveness at holiday-peak traffic remain unverified. Treat the scaling claims as AWS's, pending load testing.